GTC 2020: Weakly Supervised Training to Achieve 99% Accuracy for Retail Asset Protection
After clicking “Watch Now” you will be prompted to login or join.
Click “Watch Now” to login or join the NVIDIA Developer Program.
Weakly Supervised Training to Achieve 99% Accuracy for Retail Asset Protection
Bhavesh Patel , Dell EMC | Matt Scott, Malong Technologies
We'll present an efficient way to train deep neural networks on large-scale, weakly-supervised data learned from regular retail customer asset-protection behaviors, without any expert annotation. We'll develop a principled learning strategy by leveraging curriculum learning to effectively handle a massive number of noisy labels and data imbalance. Our new learning curriculum measures the complexity of data using its distribution density in a feature space, and ranks that complexity without supervision. This allows for an efficient implementation of curriculum learning on large-scale retail images, resulting in a high-performance convolutional neural network model, where the negative impact of noisy labels is reduced substantially.